ByteDance announced five new models at the Volcano Engine FORCE conference today, with Seedance 2.5 being the most notable.
Core Capabilities
Seedance 2.5's biggest breakthrough is generating single video clips up to 30 seconds long without any post-stitching. The videos can include scene changes and tempo shifts within a single take. Previously, most models started showing visual drift or style inconsistency beyond 10 seconds.
The model can process up to 50 additional inputs simultaneously, including reference images and audio. For multi-character scenes, each character can have its own reference image to maintain consistency. Generated videos can also be re-edited afterward while preserving the visual style.
Seedance 2.0 Update
Seedance 2.0 now supports native 4K resolution with 10-bit color depth, significantly improving color representation.
Other Models Released
ByteDance released five models total at the FORCE conference. Besides Seedance 2.5 and the Seedance 2.0 update, there are three more:
- Doubao 2.1 Pro: A language model that costs about 80% less than Claude Opus 4.6, according to ByteDance
- Seedream 5.0 Pro: An image model
- Seed-Audio 1.0: An audio model
Availability
Seedance 2.5 is expected to launch in early July through Volcano Engine's enterprise service platform. Individual users can try it through Jimeng AI (jimeng.jianying.com).
Competition
The AI video space is heating up with Google's Veo3, OpenAI's Sora, and Kuaishou's Kling. ByteDance's 30-second single-take capability and 50-input processing are technically impressive on paper, but real-world performance will need user validation after launch.